Technical Summary: Binary Road Surface Classification Using Machine Learning on Production Vehicle Signals During Cruising

Problem Statement

Accurate real-time estimation of tire-road friction is critical for vehicle safety systems, including collision warning, adaptive cruise control, and autonomous emergency braking (AEB). Current AEB systems often assume a dry surface friction value to avoid false positives, which limits their ability to react early on slippery surfaces like snow and ice.

Traditional methods for estimating friction rely on dynamics-based recursive estimators (e.g., Recursive Least Squares) that calculate the slip slope to infer tire stiffness. However, these methods face significant limitations:

1. Dependency on Slip: They require micro-slip to function effectively. During cruising conditions where slip is minimal or non-existent, these estimators become ineffective.
2. Sensor Limitations: Production-grade sensors (such as wheel speed sensors) struggle to measure or accurately estimate the micro-slip necessary to distinguish between different road surfaces in low-excitation scenarios.
3. Environmental Variability: Factors such as temperature, tire wear, and inflation pressure complicate the correlation between tire stiffness and grip levels.

The core challenge addressed in this work is the inability to reliably classify road surface conditions (specifically distinguishing between high-grip and low-grip states) using only onboard vehicle signals during low-excitation cruising.

Methodology

The authors propose a data-driven framework to correlate vehicle dynamics behavior with road surface conditions. The study utilizes a dataset comprising approximately 10 hours of snow/ice data and 20 hours of dry/damp data collected from an all-wheel-drive electric vehicle equipped with winter tires. Data was sampled at 100 Hz.

Data Pre-processing and Labeling

• Labeling: An internal optical sensor was used to automatically label time-series buffers as "grip" (dry/moist) or "slip" (snow/ice). Wet conditions were excluded due to the difficulty in measuring water film thickness and the similar tire behavior of dry and wet surfaces in the linear regime.
• Normalization: Global min-max normalization was applied to bound signals between 0 and 1, preserving relative magnitude differences essential for time-series analysis.
• Windowing: A sliding window approach was used to batch short buffers of vehicle signals.

Two Complementary Approaches

The study evaluates two distinct machine learning architectures:

1. Feature-Based XGBoost Model

• Input Features: Selected based on literature for detecting slipperiness during cruising:
  • Ratio of driven vs. non-driven wheel speeds.
  • Root Mean Square (RMS) of slip ratio.
  • RMS of slip ratio derivative.
  • RMS of longitudinal acceleration.
• Constraints: The dataset was constrained to 4-second buffers where the median longitudinal acceleration was less than 0.05G, wheel speed differences were under 2 kph, and steering/yaw rates were within specific low-excitation limits.
• Algorithm: An ensemble method based on Extreme Gradient Boosting (XGBoost) was used to learn non-linear discriminative functions from the selected features.

2. End-to-End 1D Convolutional Neural Network (CNN)

• Input: Multi-channel time-series buffers (2 seconds duration, 0.1s stride) containing 13 channels: wheel speeds, wheel torques, longitudinal/lateral/vertical accelerations, steering angle, and yaw rate.
• Architecture: A 1D CNN designed to learn time-based and implicit frequency-based pattern differences directly from raw data, avoiding the computational cost of explicit frequency domain transformations.
• Constraints: Buffers were constrained to median acceleration < 0.1G, wheel speed differences < 5 kph, and wider steering/yaw limits compared to the feature-based model.

Key Results

Feature-Based Model Performance

• Test Set Metrics: Achieved 98% recall for the grip class and 92% recall for the slip class. Precision was 92.5% for grip and 97.9% for slip.
• Feature Importance: The "Slip ratio derivative RMS" was the most important feature (score: 0.4059), followed by the "Ratio of driven/non-driven speeds" (0.2748).
• Observation: While individual features showed minimal variation between classes during cruising, the ensemble method successfully learned discriminative interaction terms.

1D CNN Model Performance

• Test Set Metrics: Achieved 98.85% recall for the grip class and 89.03% recall for the slip class. Precision was 90.0% for grip and 98.7% for slip.
• Comparison: The CNN was more conservative in predicting the slip class compared to the tree-based model.
• Interpretability (GradCAM): The CNN relied heavily on acceleration channels (Longitudinal: 0.2654, Lateral: 0.1224). The authors note that IMU acceleration data can be noisy (due to pavement quality, engine noise, chassis vibrations), suggesting a risk of overfitting to noise rather than physical grip signatures.

Real-World and Embedded Testing

• Generalization: Both models demonstrated the ability to detect snow on unseen test runs with minimal false positives on dry surfaces. The tree-based model showed slightly better snow detection but had a narrower operational domain. The CNN operated in a wider range of conditions but showed sensitivity to false positives during acceleration/braking events.
• Real-Time Feasibility: Both algorithms were implemented on Raspberry Pi hardware, achieving inference times under 100 milliseconds. A smoothing logic using a circular buffer was employed to stabilize outputs, resulting in an estimated detection delay of approximately 3 seconds.

Significance and Claims

The paper claims to demonstrate that grip/slip road-state classification during low-excitation cruising can be performed using only production-feasible onboard vehicle signals.

Key contributions and limitations highlighted by the authors include:

• Feasibility: Data-driven methods offer a viable alternative to traditional physics-based estimators when slip is insufficient for dynamic calculations.
• Scalability: Unlike traditional methods that may be vehicle-specific, data-driven models have the potential to scale across different vehicle and tire platforms.
• Modest Claims: The authors explicitly state that this is an early step. They acknowledge several limitations:
  • The problem is simplified to binary classification (grip vs. slip), excluding wet, slush, and surface transitions.
  • Evaluation on unseen drives is qualitative; specific metrics like detection delay in split-mu scenarios and robustness across diverse pavement types are not yet quantified.
  • The train/test split strategy involving sliding windows may introduce temporal overlap (data leakage), requiring more rigorous examination.
  • Production integration aspects and ASPICE compliance are not addressed.

The authors conclude that while promising, future work must extend to multi-class classification, quantify real-time performance metrics in complex scenarios (e.g., jump-mu), and validate robustness across diverse driving conditions before these methods can be certified for safety-critical applications.